auto-triage

Introduction

This project is a reproduction of the SIGGRAPH 2016 paper Automatic Triage for a Photo Series written in Python with the help of Keras and TensorFlow.

Structure

data/                # benchmark dataset (Princeton Adobe photo triage dataset)
| demos/             # three demo scenarios (jpeg files)
| download.sh        # data downloading and preparation
src/                 # source code
| data.py            # data loading and preprocessing
| models.py          # models with different settings
| train.py           # script for training
| evaluate.py        # script for evaluation
| predict.py         # script for prediction

Usages

Requirements

• Python 2.7
• OpenCV 2
• Keras 2.0+
• TensorFlow 1.0+

Preparations

cd data/ && sh ./download.sh

Training

cd src/ && python train.py <options>

Options

--exp                experiment identifier (default: default)
--gpu                GPU used for training (default: 0)
--epochs             number of training epochs (default: 16)
--batch              mini-batch size (default: 4)
--model              model (default: vgg16)                          (vgg16 | vgg19 | resnet50)
--siamese            weight sharing (default: share)                 (share | separate)
--weights            transfer learning (default: imagenet)           (imagenet | random)
--module             feature interaction (default: subtract)         (subtract | bilinear | neural)
--activation         activation function (default: tanh)             (tanh | relu)
--regularizer        regularizatiation function (default: l2)        (l2 | none)

Evaluation

cd src/ && python evaluate.py <options>

Options

--exp                experiment identifier (default: default)
--gpu                GPU used for evaluation (default: 0)

Prediction

cd src/ && python predict.py <options> <image-list>

Options

--exp                experiment identifier (default: default)
--gpu                GPU used for prediction (default: 0)

Examples

In order to produce the prediction for demo scenario 1, you may use the following command:

cd src/ && python predict.py ../data/demos/scenario-1/scenario-1-a.jpg ../data/demos/scenario-1/scenario-1-b.jpg

Also, you may use the following command for short:

cd src/ && python predict.py ../data/demos/scenario-1

License

This project is released under the open-source MIT license.